The Skyline Operator Borzsonyi S, Kossmann D, Stocker K. Classic Paper for Reference Noted by David Hsu

Content Introduction SQL Extensions Implementation of the Skyline Operator Other Skyline Algorithms Performance Experiments and Results Related Work Conclusion

For a better and general understanding about Skyline ， these Chinese papers is highly recommended. [1] 朱琳，关佶红, 周水庚.Skyline 计算综述 [J]. 计算机工程与应用， 2008 [2] 魏小娟，杨 婧，李翠平，陈 红.Skyline 查询处理 [J]. 软件学报， 2008,6

Introduction The database system at your travel agents’ is unable to decide which hotel is best for you, but it can at least present you all interesting hotels. Interesting are all hotels that are not worse than any other hotel in both dimensions. We call this set of interesting hotels the Skyline. From the Skyline, you can now make your final decision, thereby weighing your personal preferences for price and distance to the beach.

Introduction : Example Computing the Skyline is known as the maximum vector problem [KLP75,PS85].We use the term Skyline because of its graphical representation (see below). More formally, the Skyline is defined as those points which are not dominated by any other point. A point dominates another point if it is as good or better in all dimensions and better in at least one dimension. For example, a hotel with price=$50 and distance = 0.8 miles dominates a hotel with price=$100 and distance = 1.0 miles.

Introduction : nice properties One of the nice properties of the Skyline of a set is that for any monotone scoring function,if maximizes that scoring function, then is in the Skyline. In other words, no matter how you weigh your personal preferences towards price and distance of hotels, you will find your favorite hotel in the Skyline. In addition, for every point in the Skyline, there exists a monotone scoring function such that maximizes that scoring function. In other words, the Skyline does not contain any hotels which are nobody’s favorite.

由 skyline 的定义, 得到如下的重要性质 : 性质 1 如果 p 点支配 q 点, 任何一个在所有维上都单调的函数对 p 点的打分都优于 对 q 点的打分。 性质 2 任何一个在所有维上都单调的函数, 如果在数据库的点 p 上取得了所有点 中的函数最大值, 那么 p 点一定是此数据库的 skyline 点。 反之也成立 : 性质 3 对数据库中的任何一个 skyline 点 p, 必然存在一个在所有维上都单调的函 数, 使得 p 在此函数上取得的值是数据库所有点中取得的函数最大值。 Introduction

In this work, we show how the Skyline operation can be integrated into a database system. We will first describe possible SQL extensions in order to specify Skyline queries. Then, we will present and evaluate alternative algorithms to compute the Skyline; it will become clear that the original algorithm of [KLP75, PS85] has terrible performance in the database context. We will also briefly discuss how standard index structures such as B-trees and R-trees can be exploited to evaluate Skyline queries. In addition, we will show how the Skyline operation can interact with other query operations; e.g., joins, group-by, and Top N. At the end, we will discuss related work, give conclusions, and make suggestions for future work. Introduction : organization of this paper

SQL Extensions d1,...,dm denote the dimensions of the Skyline ; e.g., price, distance to the beach, or rating. MIN,MAX, and DIFF specify whether the value in that dimension should be minimized, maximized, or simply be different. For example, the price of a hotel should be minimized (MIN annotation) whereas the rating should be maximized ( MAX annotation). In our Skyline of Manhattan example, two buildings that have different x coordinates can both be seen and therefore both may be part of the skyline; as a result, the x dimension is listed in the SKYLINE OF clause of that query with a DIFF annotation. The optional DISTINCT specifies how to deal with duplicates (described below).

SQL Extensions The semantics of the SKYLINE OF clause are straightforward. The SKYLINE OF clause is executed after the SELECT...FROM...WHERE...GROUP BY...HAVING...part of the query, but before the ORDER BY clause and possibly other clauses that follow (e.g., STOP AFTER for Top N[CK97]). The SKYLINE OF clause selects all interesting tuples; i.e., tuples which are not dominated by any other tuple. Extending our definition from the introduction tuple p=(p1,...,pk,pk+1,...,pl,pl+1,...,pm,pm+1,...,pn) dominates tuple q=(q1,...,qk,qk+1,...,ql,ql+1,...,qm,qm+1,...,qn) for a Skyline query

SQL Extensions If pi = qi for all i=1,...,m, then p and q are incomparable and may both be part of the Skyline if no DISTINCT is specified. With DISTINCT, either p or q are retained(the choice of which is unspecified). The values of the attributes dm+1,..., dn are irrelevant for the Skyline computation, but these attributes are of course part of the tuples of the Skyline(i.e., there is no implicit projection).Note that it does not matter in which order the dimensions are specified in the SKYLINE OF clause; for ease of presentation, we put the MIN dimensions first and the DIFF dimensions last. In addition, a one-dimensional Skyline is equivalent to a min, max, or distinct SQL query without a SKYLINE OF clause.

SQL Extensions Qerry1: cheap hotels near the beach in Nassau and the Skyline of Manhattan. Query3: salespersons who were very successful in 1999 and have low salary; these people might be eligible for a raise. Query4: cheap hotels near the beach in Nassau ; this time, however, at most two hotels are returned because the query specifies price categories ; within each price category the user is only interested for the hotel with the smallest distance to the beach. Naturally, attributes which are specified in the SKYLINE OF clause may also be used in any other clause. To eliminate outrageously expensive hotels, for example, the WHERE clause of the first query of Figure could be extended by a price predicate.

Implementation of the Skyline Operator Our approach is to extend an existing (relational, object-oriented or object relational ) database system with a new logical operator that we refer to as the Skyline operator. The Skyline operator encapsulates the implementation of the SKYLINE OF clause. The implementation of other operators (e.g., join ) need not be changed. According to the semantics of Skyline queries, the Skyline operator is typically executed after scan, join, and group- by operators and before a final sort operator, if the query has an ORDER BY clause.

Implementation of the Skyline Operator Just like join and most other logical operators, there are several different (physical) ways to implement the Skyline operator. In this section, we will describe seven variants: three variants based on a block-nested-loops algorithm; three variants based on divide-and-conquer; and one special variant for two dimensional Skylines. Furthermore, we will show how Skyline queries can be implemented on top of a relational database system, without changing the database system at all; it will become clear, however, that this approach performs very poorly.

Implementation: Translating a Skyline Query into a Nested SQL Query ※ Essentially,this approach corresponds to the naive “nested-loops” way to compute the Skyline because this query cannot be unnested [ GKG+97, BCK98 ]; as we will see in the following subsections, we can do much better. ※ If the Skyline query involves a join or group-by(e.g., the third query of Figure 3 ),this join or group-by would have to be executed as part of the outer query and as part of the subquery. ※ As we will see in Section 4 the Skyline operation can be combined with other operations(e.g., join or Top N)in certain cases, resulting in little additional cost to compute the Skyline.

Implementation: Two-dimensional Skyline Operator A two-dimensional Skyline can be computed by sorting the data. If the data is topologically sorted according to the two attributes of the SKYLINE OF clause, the test of whether a tuple is part of the Skyline is very cheap: you simply need to compare a tuple with its predecessor. More precisely, you need to compare a tuple with the last previous tuple which is part of the Skyline. Figure 4 illustrates this approach. h2 can be eliminated because it is dominated by h1, its predecessor. Likewise, h3 can be eliminated because it is dominated by h1, its predecessor after h2 has been eliminated.

Implementation: Two-dimensional Skyline Operator Figure 5 shows why sorting does not work if the Skyline involves more than two dimensions. In this example, we are interested in hotels with a low price, a short distance to the beach, and a high rating ( many stars ). The only hotel which can be eliminated is h3: h3 is dominated by h1, but h1 is not h3’s direct predecessor. In this example, there is just one hotel between h1and h3; in general, however, there might be many hotels so that sorting does not help. There are special algorithms to deal with three- dimensional Skylines [KLP75], but for brevity we will not discuss such algorithms in this work.

Implementation: Block-nested-loops Algorithm